SD-ViT: Face Recognition ViT With Gallery-Consistent Early Exit at Any Depth
FraunhoferIGD · hf · 2026-11-10
Fraunhofer IGD released SD-ViT on Hugging Face, a training scheme for Vision Transformers where every block depth yields discriminative face embeddings.
- Training: each mini-batch randomly samples an exit depth kU{1,…,L-1}; a shared embedding head plus margin-penalty softmax loss supervises both full-depth and stochastic early exits.
- Inference: a probe can exit at any depth chosen by compute budget (edge/mobile/cloud) and remains directly comparable to a single full-depth gallery — no re-enrollment, no per-exit adapters, no architectural change.
- Takeaway: one model serves devices from tiny edge chips to cloud GPUs, cutting inference cost with minimal accuracy loss.
More from Research
- Team fine-tunes Dream-1 with just 50 examples for a USB-unplugging robot task — IsaiahBallah · 2026-11-10
- Frontier3DBench: 145 tasks test whether coding agents can replace 3D vision pipelines — Aiden_Tech_Ai · 2026-11-10
- JEPA is just a new name for old tricks: iBOT, data2vec and even CPC were basically I-JEPA — TimDarcet · 2026-11-10
- Reframe: NJU's Two-Stage Post-Training Framework Aligns Visual Agents' Observations, Beats Mix-RL and Vanilla MOPD — NJU · 2026-11-10
- Spatial Harness Grounds VLM Spatial Reasoning in Executable Physics-Engine World States — TMLR-Group-HF · 2026-11-10
- InstantHMR Runs Real-Time 3D Human Mesh Recovery on Phone NPUs: 5.5ms per Person, 12x Faster Than Peers — naturalpad · 2026-11-10